Scholarlink

<p>In recommendation scenarios and online advertisements, accurate prediction models are important for the task of click-through rate (CTR) prediction. Current prediction methods usually model higher-order feature interactions to learn complex relationships. However, during training, excessive redundant features will increase model computational complexity. Meanwhile, the process of higher-order feature interactions introduces noisy information and loses original information, thereby affecting the accurate representation of the combined features. To overcome these limitations, a CTR prediction model of fusion feature selection and layer-wise residual gated cross-network (SeLRGCN) is proposed. The SeLRGCN utilizes a feature selection layer to control the information stream in two branches, identifying important features that are jointly recognized by both information streams, which initially reduces the complexity of higher-order interactions. On this basis, a layer-wise residual gated cross-network is designed. The network reinforces feature combinations through considering the guiding role of original combined features to mitigate premature loss of original information. Meanwhile, by designing gating mechanism to filter out ineffective feature interactions, it mitigates the impact of noisy information on the model. SeLRGCN is trained on large-scale real-world datasets such as Criteo, Avazu and MovieLens. The depth of architecture and parametric richness enable stable modeling of sophisticated feature cross-correlations at high orders. Computational results on three public datasets indicate that SeLRGCN model achieves effective improvement in prediction performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Click-through rate prediction via feature selection and layer-wise residual gated cross-network

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Yuan Zhang,
  • Tiantian Sheng,
  • Yaqin Fan,
  • Lunping Duan

摘要

In recommendation scenarios and online advertisements, accurate prediction models are important for the task of click-through rate (CTR) prediction. Current prediction methods usually model higher-order feature interactions to learn complex relationships. However, during training, excessive redundant features will increase model computational complexity. Meanwhile, the process of higher-order feature interactions introduces noisy information and loses original information, thereby affecting the accurate representation of the combined features. To overcome these limitations, a CTR prediction model of fusion feature selection and layer-wise residual gated cross-network (SeLRGCN) is proposed. The SeLRGCN utilizes a feature selection layer to control the information stream in two branches, identifying important features that are jointly recognized by both information streams, which initially reduces the complexity of higher-order interactions. On this basis, a layer-wise residual gated cross-network is designed. The network reinforces feature combinations through considering the guiding role of original combined features to mitigate premature loss of original information. Meanwhile, by designing gating mechanism to filter out ineffective feature interactions, it mitigates the impact of noisy information on the model. SeLRGCN is trained on large-scale real-world datasets such as Criteo, Avazu and MovieLens. The depth of architecture and parametric richness enable stable modeling of sophisticated feature cross-correlations at high orders. Computational results on three public datasets indicate that SeLRGCN model achieves effective improvement in prediction performance.